Periodic Reporting for period 2 - Raven (AI-based autonomous flight control for the aircraft of today and electric VTOLs of tomorrow)
Período documentado: 2020-11-01 hasta 2021-10-31
Autonomous flight has the potential to address 3 important challenges that aviation industry is facing today & will face in the nearest future:
1.Despite the stellar safety record of aviation, pilots are humans & they make mistakes—75% of general aviation accidents in the US are caused by human error
2.Piloting is a very complex activity & requires a lot of training
3.There is already a shortage of qualified pilots globally & it will only get worse
Raven is an airframe-agnostic solution targeted at fly-by-wire aircraft of today (fixed-wing airplanes & helicopters) & tomorrow (eVTOLs & others). It is a combination of custom-designed neural networks, core avionics software, computer vision algorithms & special-purpose aviation-grade hardware.
We are developing a special kind of neural networks that are deterministic by design, built specifically for the purposes of certification in aviation. They learn while they are getting trained, they don’t learn when they are piloting aircraft. They run in the dedicated environment & on certifiable hardware that we are creating. This is a major innovation over the generic state-of-the-art in AI.
The main objective of this EIC Accelerator H2020 project was to industrialize Raven so that it performs better than a human pilot on any measurable dimension and could pass the certification tests currently devised for human pilots.
1.Despite the stellar safety record of aviation, pilots are humans & they make mistakes—75% of general aviation accidents in the US are caused by human error
2.Piloting is a very complex activity & requires a lot of training
3.There is already a shortage of qualified pilots globally & it will only get worse
Raven is an airframe-agnostic solution targeted at fly-by-wire aircraft of today (fixed-wing airplanes & helicopters) & tomorrow (eVTOLs & others). It is a combination of custom-designed neural networks, core avionics software, computer vision algorithms & special-purpose aviation-grade hardware.
We are developing a special kind of neural networks that are deterministic by design, built specifically for the purposes of certification in aviation. They learn while they are getting trained, they don’t learn when they are piloting aircraft. They run in the dedicated environment & on certifiable hardware that we are creating. This is a major innovation over the generic state-of-the-art in AI.
The main objective of this EIC Accelerator H2020 project was to industrialize Raven so that it performs better than a human pilot on any measurable dimension and could pass the certification tests currently devised for human pilots.
We have carried out the following activities:
- We have completed the creation of a simulation engine. In the course of the project we used 625 generated scenarios (encounters) and could reproduce flights in simulation, including augmented sets, as needed. We completed 340 hours of real test flights on fixed wing and helicopters (also including representative trajectories for eVTOLs and Urban Mobility use case) what brought us to the total of 200 TB of data collected. We prepared the infrastructure to train and experiment with training neural networks giving us scale and flexibility needed. The Spotter neural networks that had been trained on synthetic data generated in simulation environments were verified against real-life footage, and, additionally, in real-time and onboard an aircraft.
- We have successfully specified and tested the various camera components, which comprise the camera design for Raven. We selected the necessary components (CPUs, GPUs, peripherals, etc.) to design and build a ruggedised system that can execute in real time Raven neural networks. The selected components were integrated into a custom enclosure, connected to several cameras that we built and several mounting options were prepared for the trial cases.
- We managed to complete a trial case with eVTOL aircraft.Three trial case completed with external partners on time. Ten systems deployed in flight with customers and for demonstrations, and three on the ground as test stations.
- We have been ensuring the compliance of our system with the certification requirements. We have also collaborated with EASA (The European Union Aviation Safety Agency ) mainly investigating the existing regulation standards, and major reports on the use of machine learning-based systems in safety-critical systems.
- We have updated and further developed the Communication Strategy containing the objectives and action list specifically on publishing and promoting relevant project results, attracting the target audiences’ attention to Raven. We have successfully developed our social media accounts, launched a new website and also participated in industry-related events.
During the second year of the project, our team has continued to proceed with the processes set up in the first period:
- Communication with the EC has been conducted via project dedicated email,
- Project finances were monitored
- Regular meetings with the WP Leaders were held on weekly basis,
- Risks were monitored on a regular, monthly basis and we have stayed vigilant to the inherent risks of the project and continuously worked on risk mitigation,
- Project documentation has been maintained in the project documentation repository especially related to deliverable, periodic reports, procedures and resource allocation
- We have completed the creation of a simulation engine. In the course of the project we used 625 generated scenarios (encounters) and could reproduce flights in simulation, including augmented sets, as needed. We completed 340 hours of real test flights on fixed wing and helicopters (also including representative trajectories for eVTOLs and Urban Mobility use case) what brought us to the total of 200 TB of data collected. We prepared the infrastructure to train and experiment with training neural networks giving us scale and flexibility needed. The Spotter neural networks that had been trained on synthetic data generated in simulation environments were verified against real-life footage, and, additionally, in real-time and onboard an aircraft.
- We have successfully specified and tested the various camera components, which comprise the camera design for Raven. We selected the necessary components (CPUs, GPUs, peripherals, etc.) to design and build a ruggedised system that can execute in real time Raven neural networks. The selected components were integrated into a custom enclosure, connected to several cameras that we built and several mounting options were prepared for the trial cases.
- We managed to complete a trial case with eVTOL aircraft.Three trial case completed with external partners on time. Ten systems deployed in flight with customers and for demonstrations, and three on the ground as test stations.
- We have been ensuring the compliance of our system with the certification requirements. We have also collaborated with EASA (The European Union Aviation Safety Agency ) mainly investigating the existing regulation standards, and major reports on the use of machine learning-based systems in safety-critical systems.
- We have updated and further developed the Communication Strategy containing the objectives and action list specifically on publishing and promoting relevant project results, attracting the target audiences’ attention to Raven. We have successfully developed our social media accounts, launched a new website and also participated in industry-related events.
During the second year of the project, our team has continued to proceed with the processes set up in the first period:
- Communication with the EC has been conducted via project dedicated email,
- Project finances were monitored
- Regular meetings with the WP Leaders were held on weekly basis,
- Risks were monitored on a regular, monthly basis and we have stayed vigilant to the inherent risks of the project and continuously worked on risk mitigation,
- Project documentation has been maintained in the project documentation repository especially related to deliverable, periodic reports, procedures and resource allocation
We are developing Raven to augment pilots in aircraft cockpits, reduce their workload & further improve aircraft safety. We want the Raven components to extend pilot capabilities with an additional degree of flight automation & provide pilots with high quality advice during critical & emergency situations. Raven is the next logical step in flight automation & reduction of labour in the cockpits. This will result in a fully autonomous passenger eVTOL. The need for reducing pilot workload & resulting fatigue will become even more apparent as new generation planes start flying ultra-long routes. We are developing a system that can show the same holistic awareness of all flight-related issues as a human pilot.
With Raven, we will be able to make the pilots even more efficient. With Raven executing flights autonomously, an experienced pilot can be remotely supervising dozens of flights at the same time, interfering only when required.
This news article on Daedalean summarises our progress, including project RAVEN, very well: https://evtol.news/news/swiss-startup-artificial-intelligence-uam(se abrirá en una nueva ventana)
With Raven, we will be able to make the pilots even more efficient. With Raven executing flights autonomously, an experienced pilot can be remotely supervising dozens of flights at the same time, interfering only when required.
This news article on Daedalean summarises our progress, including project RAVEN, very well: https://evtol.news/news/swiss-startup-artificial-intelligence-uam(se abrirá en una nueva ventana)